.png){kind=icon}
.png){kind=icon}
Return
Embedding-Only Uplink for Onboard Retrieval Under Shift in Remote Sensing
.png){kind=icon}
PRE
AI
S
DOI：10.1109/lgrs.2026.3718430.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Downlink bandwidth forces satellites to triage observations before transmitting raw imagery, yet onboard storage of full image databases is impractical. We study a strict embedding-only uplink regime: the ground station sends only compact embeddings plus metadata, the onboard system stores them in a LanceDB vector index, and each new capture is answered by a task-appropriate decision head over the retrieved neighborhood, emitting under 1 kB of telemetry per query. On a scaled multitask benchmark over Sentinel-2 L2A (hazard, cross-time change, and cloud cover) and SpaceNet-2 RGB buildings, within each sensor/encoder regime (Sentinel-2 + OlmoEarth; SpaceNet-2 RGB + SigLIP), the same uplinked embedding table drives every head we evaluate; the optimal head is selected by task structure rather than difficulty—centroids dominate the discrete cross-time change task, the linear probe edges out <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>NN on continuous cloud cover, and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$k$ </tex-math></inline-formula>NN ties with centroids on cross-AOI buildings (under the ten-seed resampling protocol, all retrieval-versus-random primary contrasts are significant and the head-vs-head contrasts above are significant or tied under paired Wilcoxon tests; these tests quantify resampling variance over fixed scene pools, not population-level generalization). Our analyses span: 1) an encoder-size sweep showing OlmoEarth-Nano matches Base on clean inputs at a much smaller uplink; 2) an INT8 quantization study at no measurable retrieval-metric loss; 3) a cross-season audit showing absolute cosine cannot separate the same-location from out-of-location queries—rank-based retrieval still works, but an absolute-cosine abstain threshold does not transfer and must be recalibrated; and 4) an end-to-end OlmoEarth-Nano latency measurement on a Jetson AGX Orin (13.9 ms <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\mathrm {p_{50}}$ </tex-math></inline-formula>) supporting onboard-compute plausibility on an edge-class COTS surrogate, subject to flight qualification. The encoder and quantization savings compose to a roughly <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$24\times $ </tex-math></inline-formula> smaller uplink relative to an FP32 Base baseline in the low-noise regime; under heavy sensor noise Base retains a large accuracy margin, so the encoder choice is conditioned on imager noise.
Keywords:
Distribution shift
edge inference
embedding compression
onboard processing
Sentinel-2
vector retrieval
Journal
I
IF:
4.4
Papers:
486
Citations:
0
Organization
Cited Papers